[Inference Providers] isolate image-to-image payload build for HF Inference API

[hanouticelina]

This PR is a prerequisite for #1427.

It refactors the payload construction for hf-inference by isolating it into a separate async function. Note that adding a new async function to build the payload is necessary because HFInferenceImageToImageTask.preparePayload cannot be made async, yet the payload construction requires asynchronous operations. A similar pattern has already been implemented for automaticSpeechRecognition with fal.

[Wauplin]

(high level feedback)

[SBrandeis]

HFInferenceImageToImageTask.preparePayload cannot be made async

I don't understand - why can't it be made async?

[zeke]

Hey folks. Looking forward to getting this shipped to unblock #1427 -- anything I can do to help?

[SBrandeis]

Ok to merge to unlock downstream tasks

I'm not a fan of having 2 different functions for the same purpose, and I advocate to make preparePayload async instead

I also understand it's a bug and breaking API change, so let's merge as is.

[Wauplin]

I don't understand - why can't it be made async?

Mainly because we want makeRequestOptionsFromResolvedModel to be a sync method to be callable in inference snippets for instance. Kinda a weak reason but on the other hand, breaking this would be quite painful. Agree this solution is not ideal but I can't think of a better one rn